Neural Circuit Inference from Function to Structure Esteban Real, Hiroki Asari, Tim Gollisch, Markus Meister  Current Biology  Volume 27, Issue 2, Pages 189-198 (January 2017) DOI: 10.1016/j.cub.2016.11.040 Copyright © 2017 Elsevier Ltd Terms and Conditions

Figure 1 A Progression of Circuit Models Constrained by Retinal Anatomy (A) Schematic of the circuit upstream of a ganglion cell in the vertebrate retina. Photoreceptors (P) transduce the visual stimulus into electrical signals that propagate through bipolar cells (B) to the ganglion cell (G). At both synaptic stages, one finds both convergence and divergence, as well as lateral signal flow carried by horizontal (H) and amacrine (A) cells. The bipolar cell and its upstream circuitry are modeled by a spatiotemporal filter, a nonlinearity, and feedback (bipolar cell module [BCM]; blue). The amacrine cell introduces a delay in lateral propagation (amacrine cell module [ACM]; red). The ganglion cell was modeled by a weighted summation, another nonlinearity, and a second feedback function (ganglion cell module [GCM]; green). Drawings after Polyak, 1941. (B) LN model. A different temporal filter is applied to the history of each bar in the stimulus. The outputs of all of these filters are summed over space. The resulting signal is passed through an instantaneous nonlinearity. (C) LNSN model. The stimulus is first processed by partially overlapping, identical BCMs, each of which consists of its own spatiotemporal filter and nonlinearity. Their outputs are weighted and summed by the GCM, which then applies another instantaneous nonlinearity to give the model’s output. For display purpose, the BCMs are shown here to span only three stimulus bars, but they spanned seven bars in the computations. (D) LNSNF model. This is identical to the previous one, except that the GCM (depicted here) has an additional feedback loop around its nonlinearity. (E) LNFSNF model. This is identical to the previous one, except that the BCMs (one of which is depicted here) have an additional feedback loop around their nonlinearities. This new feedback function is the same for all BCMs. (F) LNFDSNF model. This is identical to the previous one, except that there is a delay inserted between each BCM and the GCM. These delays are allowed to vary independently for each BCM. (G) A count of the free parameters in the LNFDSNF model, color coded as in the model diagram. Except for the total (108), the numbers here also apply to the LNSN, LNSNF, and LNFSNF models. The LN model has 186 free parameters in the linear filter (31 spatial positions, each with six-parameter temporal filter as in Equations S3–S5) and one in the nonlinearity. See also Figures S1 and S3. Current Biology 2017 27, 189-198DOI: (10.1016/j.cub.2016.11.040) Copyright © 2017 Elsevier Ltd Terms and Conditions

Figure 2 The High Precision of Retinal Responses Allows a Sensitive Discrimination of Circuit Models (A and B) Response of a sample ganglion cell to repetitions of the stimulus (A; zoom-in to one of the firing epoch in B). (Top) Each row in the raster denotes spikes from a single stimulus repeat. (Bottom) The time course of the firing rate (black; SE in gray) and that of the output of the models fitted to the same cell (blue, LN model; red, LNFDSNF model) are shown. See also Figure S2. (C and D) A performance summary of all models reveals the most effective circuit features. The example cell in (A) and (B) is highlighted in orange. (C) Explained variance (EV) of individual cells (gray line for each cell) across models (distinct colors) is shown. LN, 0.25 ± 0.15; LNSN, 0.29 ± 0.15; LNSNF, 0.38 ± 0.15; LNFSNF, 0.40 ± 0.18; LNFDSNF, 0.42 ± 0.16; median (black horizontal bar) ± interquartile range. (D) Variance explained by each model plotted as a ratio to the variance explained by the LN model is shown. Each point along the horizontal axis corresponds to a different ganglion cell, and they are sorted based on their visual response type and ordered by increasing variance ratio under the most complex model. Note the substantial jump in performance from introducing a nonlinearity at the bipolar cell output (blue to indigo) and from introducing feedback (indigo to green). See also Figure S7. Current Biology 2017 27, 189-198DOI: (10.1016/j.cub.2016.11.040) Copyright © 2017 Elsevier Ltd Terms and Conditions

Figure 3 The LNSN Model Predicts Small Subunits of the Receptive Field Spatiotemporal filters for the BCM subunits (A) and the GCM pooling functions (B) derived from fits using the LNSN model. Results for two representative GCs are shown (left, OFF type; right, ON type), whose spatiotemporal receptive fields are shown in (C). All panels have the same spatial scale. See also Figures S5A and S5B. Current Biology 2017 27, 189-198DOI: (10.1016/j.cub.2016.11.040) Copyright © 2017 Elsevier Ltd Terms and Conditions

Figure 4 The LNSN Model Predicts a Diversity of Transfer Functions at the Bipolar Cell Synapse The internal nonlinearity of the BCM module inferred by the LNSN circuit model for different ganglion cells. The horizontal axis measures the input to that nonlinearity in units of its SDs; the vertical axis shows the output of the functions. The nonlinearities are classified into three types: linear (A), monotonic nonlinear (B), and U-shaped (C). The BCM outputs are much more rectified for OFF GCs (blue) than for ON GCs (red; p = 0.005; χ2 test). See also Figure S5C. Current Biology 2017 27, 189-198DOI: (10.1016/j.cub.2016.11.040) Copyright © 2017 Elsevier Ltd Terms and Conditions

Figure 5 LNFSNF: The Importance of Feedback at the Bipolar and Ganglion Cell Level (A) Feedback kernels fitted to three representative cells, using the LNSNF model (black) and the LNFSNF model (GCM, blue; BCM, red). (B and C) The improvement from models that allow feedback is systematically related to the magnitude of the negative feedback around GCM in the LNSNF model and that around BCM in the LNFSNF model (r, correlation coefficient; p, p value for testing hypothesis of no correlation; regression line shown in case of significant correlation). Each data point shows the ratio of the E.V. values for each cell either between the LNSNF and LNSN models (B) or between the LNFSNF and LNSNF models (C) as a function of the peak negative feedback strength around BCM or GCM (colors as in A). The representative cells in (A) are highlighted in orange. See also Figures S4 and S5D. Current Biology 2017 27, 189-198DOI: (10.1016/j.cub.2016.11.040) Copyright © 2017 Elsevier Ltd Terms and Conditions

Figure 6 LNFDSNF: Time Delays from Amacrine Cell Processing Explain the Spatiotemporal Receptive Fields of Ganglion Cells (A) Delay functions (black; relative to the center) and the pooling functions (gray) for two representative cells (left, OFF type; right, ON type). The delays are longer in the surround (magenta; weighted average by the pooling weights) than in the center (green), and the transition occurs at the same spatial location where the pooling function crosses zero. (B) Population data histogram of the relative delays from the center to the surround (median value in magenta; p < 0.001; sign test). The cells in (A) are highlighted in orange. (C) Receptive fields (same cells as in A) calculated from the data (STA, top) show the surround (magenta, peak latency) lagging behind the center (green). Receptive fields calculated from the LNFDSNF model reproduce this feature (middle), but those from the LNFSNF do not (bottom). (D) The difference in the peak latency between the center and the surround across different models. Each gray line indicates a cell, and the cells in (C) are highlighted in orange. The black horizontal bars show the median values, with significant differences between the STA and those models without delays (LNSN, LNSNF, and LNFSNF models; all with p < 0.001; rank sum test). The difference in the relative delay between the STA and the LNFDSNF model is not significant (p > 0.9). Current Biology 2017 27, 189-198DOI: (10.1016/j.cub.2016.11.040) Copyright © 2017 Elsevier Ltd Terms and Conditions

Figure 7 Experimental Tests Confirm the Circuit Structure Predicted by Modeling (A) Predicted (top) and measured (middle) bipolar cell receptive fields (BC RFs), with the corresponding GC RF (bottom) obtained by a simultaneous BC-GC recording. Note that current injection into this BC significantly affected the spiking activity of this GC (Figure S6A). See also Figure S6B. (B) Spatial characteristics of the receptive fields across all BC-GC pairs with significant projections (14 GCs, each receiving projections from one of six BCs; the example in A is highlighted in orange). The full width of the receptive field center at zero crossing is significantly smaller in the predicted BC RFs (left, 243 ± 50 μm; median ± interquartile range) than in the measured GC RFs (right, 398 ± 57 μm; p < 0.001; sign test). The difference between the predicted and measured BC RFs (315 ± 68 μm) is not significant (p > 0.1). (C) The spatial profile of the pooling function of the representative GC (top, with distance from the peak in the horizontal axis) and that of the projective weight of the simultaneously recorded BC (bottom, with each dot representing the projection to a GC). See also Figure S6C. (D) Comparison between the pooling (197 ± 65 μm) and projective weights (368 ± 178 μm; median ± interquartile range of the zero-crossing radii at the excitation-inhibition transition; p = 0.01; sign test). Each gray line indicates the simultaneously recorded data (the example in C is highlighted in orange). Current Biology 2017 27, 189-198DOI: (10.1016/j.cub.2016.11.040) Copyright © 2017 Elsevier Ltd Terms and Conditions